AI Citation Volatility Statistics: How Often AI Answers and Brand Lists Change

Ask ChatGPT, Claude or Google AI for a list of brands 100 times and you will almost never see the same list twice. AI Overviews change their content 70% of the time between checks, and 40% to 60% of cited domains change within a month. Below are the sourced statistics on how often AI answers, brand lists and citations change between runs, weeks and engines, plus what that means for measurement. All sources were read on 29 October 2026.

Start the 14-day trial, no card. Track your mention rate across six engines instead of trusting one screenshot.

What are the key takeaways?

• Lists almost never repeat. There is less than a 1 in 100 chance ChatGPT or Google AI gives the same brand list in two of 100 runs, and about 1 in 1,000 of the same order (SparkToro and Gumshoe, 2,961 runs).
• Brands drop out between back-to-back runs. Only 30% of brands stayed visible from one answer to the next, and 1 in 5 held on across five runs (AirOps, 800 queries).
• Citations turn over monthly. 40.5% to 59.3% of cited domains changed in one month depending on the engine, and 70% to 90% changed over six months (Profound).
• Google's own two AI surfaces disagree. AI Overviews and AI Mode cite the same URL only 13.7% of the time, while agreeing on meaning 86% of the time (Ahrefs).
• Visibility percentage is measurable; rank position is not. Top brands still appear in 55% to 97% of repeated answers, so mention rate across many runs is a stable metric where single-run rank is noise.
• Sample size decides accuracy. A single run recovers 62.2% to 76.8% of the brands five runs would surface (Rankfor.AI study, 4,500 responses).

How often do AI answers change?

Most of the time. Ahrefs found AI Overview content changed in 70% of consecutive observations across 43,000+ keywords, with an average persistence of 2.15 days. Profound found 40% to 60% of cited domains changed within one month. Meaning stays stable: Ahrefs measured 0.95 semantic similarity between versions.

1. AI Overviews changed in 70% of consecutive observations and persisted 2.15 days on average, across 43,000+ keywords with at least 16 recorded AI Overviews each. Ahrefs notes checks were not daily, so the real rate is likely higher. Source: Ahrefs, Louise Linehan, 11 November 2025.
2. 45.5% of AI Overview citations were new on each update; only 54.5% of URLs overlapped between consecutive responses, about one URL change per re-run. Source: Ahrefs.
3. Entity overlap between consecutive AI Overviews was 54%. 37% of AI Overviews contained named entities. Source: Ahrefs.
4. Semantic consistency between versions was 0.95 out of 1.0, and search volume had no relationship with change rate (Spearman -0.014), so popular queries are not cached. Source: Ahrefs.

Our AI citation statistics page carries the broader citation-share figures; this page focuses on change over time.

Why does ChatGPT give different brand recommendations each time?

Because the answer is sampled, not looked up. Each response is generated with randomness, and web-searching engines retrieve different sources each run. SparkToro and Gumshoe found under a 1 in 100 chance of the same brand list across 100 runs. A study using content checksums found the variation was not explained by source pages changing.

5. Under 1 in 100 chance of the same brand list; about 1 in 1,000 of the same order. 600 volunteers ran 12 prompts through ChatGPT, Claude and Google AI 2,961 times, 60 to 100 runs per prompt per platform, in November and December 2025. Claude was slightly more likely to repeat a list but less likely to repeat its order. Source: SparkToro, Rand Fishkin, 28 January 2026.
6. List length itself varied from 2 or 3 items to 10 or more for the same prompt. Source: SparkToro.
7. City of Hope appeared in 69 of 71 ChatGPT answers (97%) but was the top mention in only 25. Visibility was stable; rank was not. Source: SparkToro.
8. Citation sets were identical across repeated runs only 0.01% to 0.10% of the time on Gemini, and 3% to 8% on Perplexity and SearchGPT, across 600 queries sampled daily from 3 to 11 February 2026. Page checksums showed the variation was not caused by source pages changing. Source: Sielinski, arXiv 2603.08924.
9. Real users phrase the same need very differently. 142 human-written prompts with the same intent (headphones for a travelling family member) had an average semantic similarity of 0.081. Source: SparkToro.

Four forces stack: sampling randomness in generation, different retrieved sources per run, model and product updates, and prompt wording. SparkToro's volunteers also used their own accounts and settings, so personalization adds to the spread.

How often do AI citations change week to week and month to month?

Between runs, about half of cited sources change. Over a month, 40% to 60% of cited domains change on every major engine Profound measured. Over six months, 70% to 90% change. Perplexity was the most stable engine in both Profound's monthly data and an independent run-to-run study.

10. Monthly domain drift, June to July 2025: AI Overviews 59.3%, ChatGPT 54.1%, Microsoft Copilot 53.4%, Perplexity 40.5%. About 80,000 prompts per platform were compared across 11 to 13 June and 11 to 13 July 2025. Source: Profound, 17 July 2025.
11. 70% to 90% of cited domains changed between January and July 2025 across the same engines. Source: Profound.
12. Median run-to-run Jaccard similarity of cited domains: Gemini 0.29 to 0.31, SearchGPT 0.33 to 0.40, Perplexity 0.50. A score of 1 means identical sources; 0.3 means most sources differ. Source: Sielinski, arXiv 2603.08924.
13. Only 30% of brands stayed visible from one answer to the next. Source: AirOps, 800 queries and 45,000+ citations, 23 September 2025.
14. About 57% of brands that dropped out of one response resurfaced in a later one. Brands that were both mentioned and cited were 40% more likely to resurface than brands that were only cited. Source: AirOps.
15. 28% of responses included a brand that was both mentioned and cited. Source: AirOps.

Measure	Engine	Figure	Source
Content changed between checks	Google AI Overviews	70%	Ahrefs
Citations new per update	Google AI Overviews	45.5%	Ahrefs
Domains changed in one month	AI Overviews / ChatGPT / Copilot / Perplexity	59.3% / 54.1% / 53.4% / 40.5%	Profound
Domains changed over six months	Same four	70% to 90%	Profound
Run-to-run Jaccard of sources	Gemini / SearchGPT / Perplexity	0.29 to 0.31 / 0.33 to 0.40 / 0.50	Sielinski
Brands visible in back-to-back answers	Mixed engines	30%	AirOps
Same brand list in 100 runs	ChatGPT / Google AI	under 1%	SparkToro
Shared citations, AI Overviews vs AI Mode	Google	13.7%	Ahrefs

How much do different AI engines agree with each other?

Not much on sources, more on meaning. Google's AI Overviews and AI Mode cite the same URLs only 13.7% of the time for the same query, yet reach semantically similar answers 86% of the time. Across six engines, the median buying question surfaced 38.5 organizations, 15 of them named by only one engine.

16. AI Overviews and AI Mode shared 13.7% of citations (16.3% in the top three) and 16% of words, across September 2025 US data. Only 0.51% of answer pairs were identical, and 2.51% shared a first sentence. Source: Ahrefs, AI Overviews vs AI Mode.
17. Semantic similarity between the two was 86%, with 89.7% of pairs above 0.8. A brand mentioned in an AI Overview had a 61% chance of also appearing in AI Mode. AI Mode named 3.3 entities per answer on average vs 1.3 for AI Overviews. Source: Ahrefs.
18. The median buying question surfaced 38.5 organizations across six engines, 15 of them from a single engine only. The best single engine covered a median 83.1% of the six-engine union. Source: Żatuchin, arXiv 2609.05059, 50 questions, 4,500 responses.

Our six-engine divergence index compiles the cross-engine overlap research in one place, including BrightEdge and Profound comparisons. The practical point: tracking one engine tells you little about the others.

How many runs do you need to measure AI visibility reliably?

More than one, and more for some engines than others. One run recovers 62% to 77% of the brands five runs would surface. To estimate a citation share within about plus or minus 5 points, one study needed 40 to 50 queries on Gemini, around 100 on Perplexity and at least 150 on SearchGPT.

19. A single run recovered 62.2% to 76.8% of the expected five-run brand set, depending on the engine; ten runs recovered 92.6% to 95.2% of the brands seen in 15 runs. Source: Żatuchin, arXiv 2609.05059.
20. For Perplexity Sonar source hosts, a single run recovered 59.5% of the multi-run set. Source: Żatuchin, arXiv 2609.05059.
21. Queries needed for a citation-share confidence interval of about plus or minus 5 points: Gemini 40 to 50, Perplexity about 100, SearchGPT 150 or more. SearchGPT confidence intervals spanned 3 to 6 points. Source: Sielinski, arXiv 2603.08924.
22. Two sites at 9.5% and 6.0% citation share had overlapping confidence intervals, so the apparent gap could not be called a real difference. Source: Sielinski, arXiv 2603.08924.
23. Top brands still appeared in 55% to 77% of 994 answers to 142 differently worded headphone prompts, and the top agency in a Google AI test appeared in 85 of 95 responses. Source: SparkToro.

Both arXiv papers come from authors at AI visibility vendors (IQRush and Rankfor.AI), and SparkToro's co-author works at Gumshoe, another vendor. The methods are published and the direction of every result matches, but none is peer reviewed yet.

What should you do about AI citation volatility?

Stop reporting single answers and rank positions. Track mention rate across a fixed prompt set, repeated over time and across engines, and judge changes against the noise. SparkToro's conclusion after 2,961 runs: visibility percentage across many prompts is a reasonable metric, and any tool selling "ranking position in AI" is not.

• Use percentages, not positions. City of Hope's 97% visibility was stable, yet it was not the top mention in 46 of 71 answers.
• Fix the prompt set. Changing prompts between periods mixes prompt variation with real change.
• Compare against the noise floor. With monthly domain drift of 40% to 60%, a 10-point move on a handful of prompts is not a trend.
• Watch for step changes. Product changes move everything at once; ChatGPT's clickable brand links in May 2026 changed referral patterns within a week.
• Earn mention and citation together. AirOps found brands that were both mentioned and cited were 40% more likely to resurface across runs than brands that were only cited.

For the question of how to check a single brand by hand first, see how to check if ChatGPT mentions your brand.

Measure your mention rate on a fixed prompt set across six engines. Start the 14-day trial: no card, five prompts.

What can Mentionova measure, and what can it not?

Mentionova runs your buyer prompts across six engines on a weekly, fortnightly or monthly schedule and trends mention rate and citations over time. It does not run daily, and it does not replicate each prompt 60 to 100 times per period, so treat small moves on a few prompts as noise.

• Starter is $125 a month, or $99 billed annually, for 25 prompts on three engines. The pricing page lists the larger plans.
• The 14-day trial needs no card and covers 5 prompts on all six engines.
• More prompts narrow the error bars more than more frequent runs do; the research above points to dozens of queries per engine, not a few.
• For which sources engines favour in the first place, see LLM citation statistics.

Where does each number come from?

All sources read on 29 October 2026.

• SparkToro and Gumshoe, AI brand recommendation consistency study, Rand Fishkin and Patrick O'Donnell (published 28 January 2026, updated 28 June 2026). Not peer reviewed; co-author works at Gumshoe.
• Ahrefs, AI Overviews change study, Louise Linehan and Xibeijia Guan (11 November 2025), 43,000+ keywords.
• Ahrefs, AI Overviews vs AI Mode comparison, September 2025 US data.
• Profound, AI search volatility (17 July 2025), about 80,000 prompts per platform. Vendor data.
• AirOps, citations and mentions in AI search report (23 September 2025), 800 queries and 45,000+ citations. Vendor data.
• Sielinski, "Quantifying Uncertainty in AI Visibility", arXiv 2603.08924 (2026), Perplexity, SearchGPT and Gemini. Author affiliated with IQRush.
• Å»atuchin, arXiv 2609.05059 (September 2026), 50 buying questions, six engines, 4,500 responses. Author is CEO of Rankfor.AI.
• Mentionova plan limits: mentionova.com/pricing.

What do people ask about AI citation volatility?

Why does ChatGPT give different brand recommendations each time?

Responses are sampled with randomness, web search retrieves different sources on each run, and models are updated. SparkToro found under a 1 in 100 chance of the same list across 100 runs, and about 1 in 1,000 of the same order.

How often do AI citations change?

On Google AI Overviews, 45.5% of cited URLs are new on each update, and content changes every 2.15 days on average (Ahrefs). Across engines, 40.5% to 59.3% of cited domains changed in one month (Profound).

Which AI engine is the most stable?

Perplexity, on the available data: the lowest monthly domain drift at 40.5% (Profound) and the highest run-to-run Jaccard at 0.50 (Sielinski). Gemini was least repeatable, with identical citation sets in 0.10% or fewer of runs.

Is tracking AI visibility pointless if answers change every time?

No. Rank position is noise, but mention rate is measurable. Top brands appeared in 55% to 97% of repeated answers in SparkToro's data, which is why visibility percentage across many prompts is the metric to track.

How many times should I run a prompt?

One run catches 62% to 77% of the brands five runs surface; ten runs catch 93% to 95% of a 15-run set (Żatuchin). For citation share within about 5 points, plan on 40 to 150 queries per engine (Sielinski).

Do Google AI Overviews and AI Mode cite the same sources?

Rarely. They share 13.7% of citations for the same query, though their answers are 86% semantically similar (Ahrefs). Track them as separate surfaces.

Does a drop in AI visibility mean something changed on my site?

Not necessarily. Sielinski's checksums showed citation variation was not driven by page content changes. Confirm a drop across several runs and prompts before acting on it.

Track the trend, not the screenshot. Start the 14-day trial, no card.